基于关键帧轮廓特征提取的人体动作识别方法

发布时间：2019-07-14 12:34

【摘要】：为了在人体动作识别中获得更加准确的前景分割和防止关键信息的几何丢失,提出一种利用关键帧提取关键姿势特征的人体动作识别方法。由于背景建模和差分获得的前景不准确,利用基于纹理的灰度共生矩阵提取动作轮廓,并对原图像帧进行分割;然后计算人体Blob的能量,选取最大信息内容的帧作为关键帧,关键帧的获取使得特征提取对时间的变化具有一定鲁棒性;在特征分类识别阶段,为了提高分类的准确性,提出使用支持向量机-K最近邻(support vector machine-k nearest neighbor,SVM-KNN)混合分类器完成分类。在Weizmann,KTH,Ballet和TUM 4个公开数据集上实验验证了该方法的有效性。相比于局部特征方法、全局特征方法和关键点方法等,该方法获得了更高的识别率。此外,实验结果表明,该方法在KTH和Weizmann数据集上的早期识别效果优于Ballet数据集。
[Abstract]:In order to obtain more accurate foreground segmentation and prevent the geometric loss of key information in human motion recognition, a human motion recognition method using key frames to extract key posture features is proposed. Because the prospect of background modeling and difference acquisition is not accurate, the action outline is extracted by texture-based gray co-occurrence matrix, and the original image frame is segmented. Then the energy of human body Blob is calculated, and the frame of the maximum information content is selected as the key frame, which makes the feature extraction robust to the change of time. In the stage of feature classification and recognition, in order to improve the accuracy of classification, support vector machine-K nearest neighbor (support vector machine-k nearest neighbor,SVM-KNN) hybrid classifier is proposed to complete the classification. The effectiveness of the proposed method is verified by experiments on four public datasets, Weizmann,KTH,Ballet and TUM. Compared with the local feature method, the global feature method and the key point method, this method obtains higher recognition rate. In addition, the experimental results show that the early recognition effect of this method on KTH and Weizmann datasets is better than that of Ballet datasets.
【作者单位】：平顶山学院软件学院;武汉大学计算机学院;
【基金】：国家自然科学基金项目(61503206) 河南省科技厅科技发展计划项目(142102210226)~~
【分类号】：TP391.41

【参考文献】